AI

Tyger: open-weights streaming ASR + full-duplex turn-control in one model

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 17/10/2026, 09:47:26 UTC.

Briefings

Tyger is a small open-weights model (NVIDIA Nemotron 3.5 ASR 0.6B + 14M GRU) designed for full-duplex voice agents. It operates every 80ms without lookahead, detecting turn states via microphone and agent playback. The base model runs locally with optional external expert integration. The full-stack agent (tyger-s2s) requires 28GB VRAM and delivers ~0.7s replies on RTX 5090 GPUs.

Key points

• Tyger uses a base NVIDIA Nemotron 3.5 ASR 0.6B model augmented with 14M causal GRU layers.
• The agent operates on an 80ms turn-taking interval without lookahead, distinguishing between answer, stop, and chime-in states.

Context

This model enables full-duplex voice agents with turn-control, allowing local execution without audio/text leaving the machine. The full-stack agent supports tool calls and up to 4 concurrent conversations.

Limitations

The base model is small; the full-stack agent requires 28GB VRAM. FP32 only and offline benchmarks are available; GPU testing on other models is untested.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 01:14:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicumlpz4lw5adcqmykowrobgrolqdcrnsunsotvxlldjhvvjb3p3i